ConAtt

ConAtt is an open-source PyTorch library for attention on continuous, irregularly sampled, and constrained domains. It keeps ordinary attention as a first-class case while adding explicit integration measures, geometric masks, continuous-depth dynamics, neural-operator blocks, approximation certificates, and evidence gates for fair model comparison.

The package is designed for reusable scientific and engineering work rather than a single benchmark. Initial application targets include PDE surrogate models, irregular time series, physical simulation, scientific sensing, and spatiotemporal learning.

Release status: 0.1.0a2 is a public alpha API, available from PyPI. Production stability is not yet guaranteed; scientific claims remain governed by the evidence gate below.

Install

pip install conatt[torch]

From a local wheel or in Colab before publication:

pip install /content/conatt-0.1.0a2-py3-none-any.whl

Continuous attention in three steps

import torch
from conatt.torch_continuous import ContinuousAttentionOperator1D

coordinates = torch.tensor([[0.0], [0.03], [0.20], [0.74], [1.0]])
forcing = torch.randn(8, coordinates.shape[0], 1)

model = ContinuousAttentionOperator1D(
    input_channels=1, output_channels=1, width=64, heads=4, layers=4,
    use_measure=True,
)
prediction, attention_maps = model(forcing, coordinates)

The parameter-identical ablation is obtained with use_measure=False. This makes the effect of continuous-domain normalization directly testable without changing model capacity.

Included capabilities

• NumPy and PyTorch scaled dot-product attention with positive measure weights;
• continuous 1-D operator model for irregular grids and variable resolutions;
• hybrid ConAttOperator1D coupling measure-aware attention with irregular Fourier/sine integral mixers and optional exact Dirichlet envelopes;
• LatentConAttOperator1D for measure-aware encode/process/decode between arbitrary observation points and a reusable continuous latent grid;
• fixed-step Euler, midpoint, and RK4 integration for attention vector fields;
• causal and metric-radius structure masks;
• affine physical-constraint projection with residual certificate;
• retained-mass, spectral-tail, and finite-horizon propagation certificates;
• exact propagation-aware approximation-budget allocation;
• paired-seed statistics, fairness records, and immutable experiment storage;
• deterministic Poisson data generator and controlled killer tests.
• auditable FNO and DeepONet references for local baseline checks.

Verify an installation

conatt verify
conatt benchmark-continuous --output continuous-killer-test.json

The controlled test isolates a known irregular-sampling failure. Its output is deliberately blocked from becoming a publication claim. Learned-model claims require real benchmarks, reproduced closest baselines, paired seeds, matched budgets, negative controls, and immutable artifacts.

See the five-track research program, baseline policy, and Colab quickstart.

License

MIT. See LICENSE.